What’s up with… Orange, AT&T, NBN

• Orange adopts OpenAI’s new open-weight LLMs 
• AT&T achieves first Open RAN data call on live network
• Australia’s NBN hooks up with Project Kuiper

In today’s industry news roundup: OpenAI rolls out two new open-weight language models, both of which will be deployed by Orange across 26 countries, while Qualcomm confirms on-device support from high-end Snapdragon processors; AT&T achieves Open RAN milestone with Ericsson and 1Finity in the form of a multivendor data call using third-party radios; NBN partners with Amazon Web Services’s Project Kuiper to deliver faster LEO satellite broadband to remote areas in Australia; and much more!

OpenAI has launched two new open-weight language models, gpt-oss-120b and gpt-oss-20b, designed for cost-effective and optimised deployment on consumer-grade hardware. The AI company says both models outperform similarly sized open models on reasoning tasks and tool use, and are customisable for a wide range of applications. The larger 120b model achieves near-parity with OpenAI’s o4-mini on reasoning benchmarks while running on a single 80GB GPU. The smaller 20b model is suitable for edge devices with just 16GB of memory. OpenAI says it has been working with AI Sweden, Snowflake and Orange “to learn about real-world applications of our open models”, from hosting these models on-premises for data security to fine-tuning them on specialised datasets. “A healthy open model ecosystem is one dimension to helping make AI widely accessible and beneficial for everyone,” the company noted in its announcement, adding, “broad access to these capable open-weights models created in the US helps expand democratic AI rails.” No doubt this is, in part, a response to the recent advances made by China on open large-language models (LLMs), and it is worth noting that these are OpenAI’s first open-weight models since GPT-2 in 2019. Read more.

Supporting the OpenAI product release, Orange has confirmed that it will deploy both of OpenAI’s new open-weight reasoning models across its network infrastructure, as part of a deepening collaboration between the two companies. As an early access partner, Orange says it is one of the first global operators to implement these advanced AI models, which can be flexibly deployed in Orange’s regional datacentres, on-premises servers or edge sites. This enables Orange to host AI workloads locally across its 26-country footprint, ensuring data sovereignty and compliance with regulatory requirements in Europe, the Middle East and Africa. The telco’s AI engineering team will customise and distil the OpenAI models for specific tasks, supporting innovative use cases in network operations and enterprise services, while safeguarding sensitive data. The collaboration is also central to Orange’s Responsible AI strategy, which prioritises high performance with reduced computing power, data and energy consumption. By running models within its own datacentres, Orange says it can directly manage power usage and reduce emissions. “This strategy drives new use cases to address sensitive enterprise needs, help manage our networks, enable innovating customer care solutions, including African regional languages, and much more,” said Steve Jarrett, chief AI officer at Orange. “This collaboration with OpenAI is foundational to our strategy of using state-of-the-art AI models that are both trusted and responsible.” A key focus is digital inclusion in Africa, with Orange accelerating the integration of African languages into AI models. The company plans to release customised AI models on an open-source basis to support public services in the region. Use cases include enhanced multilingual customer support, smarter network management and tailored business solutions for enterprise customers. Orange signed an agreement with OpenAI in November 2024 that gave the telco early access to OpenAI’s latest and most advanced AI models, with data processing and hosting in European datacentres.

Since OpenAI didn’t provide visual supporting material for this announcement, we thought we would simply get ChatGPT to do the job. The image above is what it came up with. Not bad… 

In related news, Qualcomm Technologies says that OpenAI’s gpt-oss-20b open-weight model is capable of running directly on flagship Snapdragon processors. It says this release marks the first time a sophisticated chain-of-thought (COT) reasoning model is available for on-device inference. Through early access and integration testing with the Qualcomm AI Engine and AI Stack, Qualcomm reported that gpt-oss-20b enables advanced reasoning entirely on-device, potentially opening the door to AI assistants that are private, low-latency and highly personalised. “We see this moment as a turning point – a glimpse into the future of AI where even rich assistant-style reasoning will be local,” wrote Qualcomm's Sachin Deshpande and Vinesh Sukumar on the company blog. They add that developers can access the model on Snapdragon-powered devices through platforms, such as Hugging Face and Ollama, with deployment resources provided via the Qualcomm AI Hub. Integration with Ollama’s lightweight LLM framework allows gpt-oss-20b to run natively on “powerful” Snapdragon devices (they don't specify which ones or provide additional benchmarks). Qualcomm expects on-device AI capabilities to grow rapidly as mobile memory increases and software becomes more efficient.

Elsewhere in telecoms-land, AT&T is ramping up the excitement over Open RAN (on which we have more to say on Friday). Only a week after announcing the successful deployment and operation of a third-party rApp on its live production network, it has now announced another Open RAN milestone. In partnership with Ericsson and 1Finity (the newly rebranded Fujitsu company), AT&T has made the first successful data call using third-party radios. The call was enabled by integrating Ericsson’s high-capacity RAN Processor 6672 (baseband) with 1Finity radios, showcasing the performance and reliability of a multivendor solution and demonstrating the practical benefits of Open RAN specifications for interoperability and flexibility in network infrastructure. Rob Soni, VP of RAN technology at AT&T, described the achievement as “a testament to AT&T’s leadership in Open RAN innovation”, highlighting the collaboration’s role in advancing network flexibility, efficiency, and creativity. “This milestone marks another significant step toward achieving our vision of an open, agile and programmable wireless network,” he added. Patrik Eriksson, vice president and head of the Mobile System Business Unit at 1Finity, described the call as “a significant validation of 1Finity’s radio expertise and the transformative potential of Open RAN,” while Johan Hultell, head of product line Cloud & Purpose-built 5G RAN at Ericsson, said the collaboration “underscores the transformative potential of robust collaboration in the Open RAN ecosystem”. For more on the Ericsson and AT&T collaboration – see What’s up with… M&A in France, Jio’s IPO, AT&T and Ericsson.

NBN, the Australian wholesale broadband provider, has announced a partnership with Amazon Web Service’s Project Kuiper to deliver high-speed, wholesale fixed broadband via low-earth orbit (LEO) satellites to regional, rural and remote parts of Australia. Project Kuiper is set to launch its service in Australia from mid-2026. Once operational, NBN plans to offer wholesale residential-grade LEO satellite broadband services to more than 300,000 premises within its existing satellite footprint, through participating retail service providers (RSPs). The agreement marks a transition from NBN’s current geostationary Sky Muster satellite services and aims to significantly enhance broadband quality and reliability for eligible customers. NBN will consult with RSPs, regional communities and stakeholders to determine speed tiers, wholesale pricing and logistics. Ellie Sweeney, CEO of NBN, described the move as “a major leap forward for customers,” adding that, “transitioning from two geostationary satellites to a constellation of LEO satellites will help to ensure the NBN network is future-ready and delivers the best possible broadband experience to customers living and working in parts of regional, rural and remote Australia.” Project Kuiper’s initial constellation will eventually include more than 3,200 satellites, with 78 already in orbit following three successful launches since April 2025. NBN’s two Sky Muster satellites are expected to remain viable and operational until approximately 2032, so the company is exploring its options regarding these assets once all existing customers have been transitioned to the new service.

Tele2 has announced the creation of the first pan-Baltic tower company through a 50/50 partnership with Global Communications Infrastructure (GCI), which is backed by Manulife Investment Management. The new company, based in Lithuania, will own around 2,700 tower and rooftop sites across Estonia, Latvia, and Lithuania. The deal values the tower company at  €560m on a debt-free basis, with Tele2 expecting cash proceeds of approximately €440m. As part of the agreement, Tele2 will be the anchor tenant under a 20-year master service agreement, ensuring long-term access to critical infrastructure for its mobile and 5G services. The parties have also committed to a 10-year investment plan aimed at building new sites across the Baltics to support expanded coverage and network growth. “We want to develop our tower assets together with a partner who brings both capital and expertise,” said Jean Marc Harion, president and CEO of Tele2. “This is a way for us to create additional value from the assets we have, together with an experienced partner who knows this business well.”  Tim Culver, CEO of GCI, added: “With a long-term perspective and commitment to investment, we see strong potential to develop high-quality, efficient and scalable infrastructure.”

– The staff, TelecomTV